VoxelForge/Standalone/Research/WorldAssemblerM3/Research48/REPORT.md

# R48 read only — who owns the tunnel mouth?

**The passage planner currently chooses the mouth.** Floating Islands then has to fit its islands and walk route around fixed tubes; caves can instead carve toward them. L140’s fixed tube layout left no valid island plan for most seeds. [CODE] `Standalone/Core/src/ResolvedLayoutInternal.h:3699–3792`; [MEASURED] `Standalone/Research/WorldAssemblerM3/Linux140/REPORT.md:15–19`

## 1. Planning order and constraints

There are two adapter paths, so there isn’t one universal order today:

- The geometry only adapter builds strata and passage configs, plans passages, then resolves authored Core geometry. [CODE] `Source/VoxelForge/Private/VoxelForgeCoreAdapter.cpp:1455–1585`
- The validated assembly path can build a manifest, including an imported Floating Islands plan, before creating the resolved layout. It copies plan anchors into that layout, then calls the passage planner. [CODE] `Source/VoxelForge/Private/VoxelForgeCoreAdapter.cpp:1621–1685, 1950–2019, 2021–2041`
- L140’s kept patch changes the Floating Islands path: first resolve the fixed passage layout, then plan islands around those passages and export the resulting anchors. The patch is not applied in the R48 source: the current adapter still calls `PlanFloatingIslandCluster` without a resolved layout. [CODE] `Standalone/Research/WorldAssemblerM3/Linux140/l140-fi-around-tunnels.patch:102–219`; `Source/VoxelForge/Private/VoxelForgeCoreAdapter.cpp:893–987`

`PlanResolvedLayoutPassagesV1` sorts the strata, uses existing route anchors where present, otherwise chooses seeded mouths, and builds passage control points with radii. Once resolved, those mouths and tubes are inputs to later planners. [CODE] `Standalone/Core/src/ResolvedLayoutInternal.h:3699–3792, 3820–3874, 4713–4819`

The generic island planner lays out seeded island centers and a route without seeing the passage layout. L133’s alternate planner bound endpoints, but did not plan against both full passage volumes. Its composed walk still had collisions and unsupported route points. [CODE] `Standalone/Core/src/FloatingIslandsInternal.h:1991–2050, 2166–2233`; historical patch `Linux133/l133-native-fi-v9.patch` in Git commit `f8cba7c:79–134`; [MEASURED] `Linux133/REPORT.md` in that commit:3–7

The cavern planner reads incoming and outgoing passages and requires them to exist, except for a Vertical Shafts basal terminal. Caves can carve through host rock to reach those mouths; L134’s default composed route passed, while L139 still failed the shell and seam checks at seed 2’s incoming mouth after three local geometry attempts. [CODE] `Standalone/Core/src/ResolvedLayoutInternal.h:2344–2359`; [MEASURED] `Linux134/REPORT.md:3–7`; `Linux139/REPORT.md:5–8, 43–52, 89–91`

## 2. Design choices

- **Content first:** each layer produces supported landing candidates and a connected approach; the passage planner chooses compatible candidates and routes to them. This can solve the fixed-height problem by letting the tunnel target viable terrain. It requires every layer type to provide reliable candidates, stable ordering and tie breaks, and passage rerouting; changed routes must still pass the existing grade, shell, seam, and ripple audits. **Size: medium to large.** [INFERENCE]

- **Negotiated landing pad:** the passage reserves a landing region and the layer guarantees support there. A point or pad contract alone is too weak: L140 checked five support probes, yet 14 seeds had no arrangement clearing the fixed tubes or could not reach both mouth heights. A useful contract must cover the capsule approach and passage exclusion envelope too. That grows toward a joint plan. **Size: medium for a pad; large for a full approach and clearance contract.** [MEASURED] `Linux140/REPORT.md:7, 15–17`; [INFERENCE]

- **Joint plan:** choose passage geometry and layer anchors together. This gives the solver room to avoid the fixed-tube infeasibility seen in L140, but costs the most: bounded deterministic search, new planner interfaces, and careful preservation of passage smoothness and audit behavior. A bounded search only guarantees feasibility within its candidate space. **Size: large.** [MEASURED] `Linux140/REPORT.md:17`; [INFERENCE]

## 3. Why ripple reaches 6–7 cm

The ripple audit samples the **composed field** passed as `Shape`, at `pose.radius − 0.18 m`, then estimates wall radius as that sample radius plus the average signed density. It cubic-detrends each 32 m window, preserving authored radius variation. Thus the result includes both the passage’s radius profile and any host-field effect on the sampled wall; it is not a passage-only measurement. [CODE] `Standalone/Source/StreamBenchmarkPipeline.h:4186–4239, 4414–4432, 4941–4944`

L67 traced a prior gauge error: omitting `pose.radius` hid an authored-radius beat. The corrected gauge found a 4.77 cm beat at 7.06 m and measured 7.27 cm on passage 1–2. L140 measured 6.22–7.19 cm on its upper passages with native islands. [MEASURED] `Linux67/REPORT.md:64–79`; `Linux140/REPORT.md:15–19`

The evidence shows that passage-radius variation alone is enough to explain a failure at this scale; it does not isolate how much native island density adds or masks. L140’s shell-preserving composition change still failed the unchanged ripple gate. [INFERENCE] `Linux140/l140-fi-around-tunnels.patch:912–924`; [MEASURED] `Linux140/REPORT.md:15–19`

## 4. Recommendation

**Use a bounded pairwise joint plan.** Have each layer planner publish certified landing candidates with capsule support and approach constraints. Let the passage planner choose compatible ports and route geometry together, then check the complete tube and composed layer before finalizing the resolved layout. Start with Floating Islands; preserve cave planners’ ability to carve to selected ports. [INFERENCE]

Keep candidate generation and tie breaking seed-stable, and finalize the layout identity only after the selected mouths, tubes, and content anchors are known. Keep the existing thresholds unchanged. [CODE] `Source/VoxelForge/Private/VoxelForgeCoreAdapter.cpp:2021–2041`; [INFERENCE]

The design passes if all 16 seeds clear the existing route, support, shell, seam, grade, mouth-profile, and 3.5 cm ripple gates, with matching Clang/GCC W1/W8/W12 hashes. Separately compare ripple on the same tube in a smooth host field and in the composed field; that isolates passage-profile ripple from host-field contribution. Kill the design if joint movement still leaves seeds infeasible, or if the passage-only profile fails ripple. [INFERENCE]

L133’s report and patch were absent from the R48 working tree; I read them from Git commit `f8cba7c`. No files were changed, and I did not build or run tests. [MEASURED]